1. Field of the Invention
The present invention relates to the transmission of data packets such as ATM packets between Local Area Networks (LAN) interconnected by a switch engine and relates in particular to a data transmission system including a self-route expandable multi-memory packet switch wherein the scheduling means is distributed.
2. Background of the Invention
Local Area Networks (LAN) such as Ethernet or token-ring networks, are generally interconnected through hubs. The hub is a system consisting of LAN adapters that communicate together through a switch card containing a switch engine. Such a switch engine can be either a shared memory switch or a crossbar switch.
The shared memory switch is a device wherein the packets received by the input ports are stored into a memory at locations the addresses of which are determined by queues containing the packet destination addresses, the packets being transmitted on the output ports as the destination addresses are dequeued. Although such a switch enables to incur a very low cell-lost rate, it presents a bottleneck due to the requirement of the memory bandwidth, the segregation of the buffer space and the centralized control of the buffer which causes the switch performance to degrade as the size of the switch increases. A traditional approach to design a large shared memory switch has been to first design a feasible size shared memory switch and then to interconnect a plurality of such modules in order to build a large switch. This general scheme of switch growth is known to cause degradation in performance of shared memory architecture as the switch grows in size insofar as the memory access controller will have to increase the number of all centralized control functions and memory operations thereby reducing drastically the access to the shared memory. A growable switch approach packet switch architecture is a plurality of shared memory switches organized in a single stage preceded by a buffer-less interconnection network. This approach does not allow global sharing of memory space along all its inputs and outputs. It is known that this approach does not provide the best buffer utilization as possible for a buffer belonging to a group of output ports to overflow under unbalanced or bursty traffic conditions.
The other technique, the crossbar switch, does not use a shared memory to store the data packets. In such a switch, the data are stored in the adapters and the switching data connection is established by sending requests to a centralized scheduler which determines whether it is possible to satisfy the requests. For this, the scheduler includes an algorithm unit which determines the best data connection to establish at each time. Such a determination is based upon the selection of the request amongst all requests received from the LAN adapters which meets some predetermined criteria such as a priority order, the selection of unicast/multicast, the selection between reserved bandwidth data and non-reserved bandwidth data, or any other criteria defined by the user.